—

Why plan the comparison instead of scoring it

Most model comparisons line up specifications in a table and declare a winner. That tells you very little about whether a model can actually carry a brief from the first frame to the final call to action. A more useful exercise is to take one realistic 30-second launch spot, the kind of deliverable a brand actually signs off on, and ask, for each model, how you would have to plan the work.

Where a model can hold the whole idea in a single generation, the plan is short. Where it cannot, the honest answer is that the workflow splits into stages, and the interesting question becomes which stage each tool is best suited to. That is the lens used here. No footage is produced; the brief is a framework, and every capability claim is drawn from each model’s documented behaviour rather than from a test render.

The hypothetical brief

Client: Vanta, a fictional wellness-technology brand. Product: the Vanta Loop, a minimalist smart ring that tracks sleep, focus, and recovery and pairs with a phone app showing a clean dashboard.

Deliverable: one 30-second launch film for paid social and the product page, cut to a horizontal master with vertical and square versions.

Recurring character: “Elena,” a thirty-something creative professional who appears in every scene. Her face, hair, and the ring on her hand must stay consistent shot to shot.

Locations: a sunrise bedroom, a morning city street, a bright daytime studio, and the same apartment at night.

Camera language: a macro product insert, a slow orbit around the ring, a handheld walking track, a static app-screen close-up, and a locked hero shot at the end.

Dialogue: Elena speaks one line to camera — “It just knows when I need to slow down.”

Music: a warm modern electronic bed that builds to a clean button under the end card.

Product shots: the ring in macro on a marble surface, the ring on Elena’s finger, the app dashboard, and the packaging with a short sonic logo at the close.

Final CTA: an end card reading “Vanta Loop. Know your rhythm. Pre-order at vanta.co,” landing on the music button and sonic logo.

Everything about this brief is deliberately ordinary. It is exactly the mix, one recurring face, several locations, real product detail, a spoken line, a music cue, and a closing CTA, that separates a model that can make a clip from a model that can complete a spot.

The reference kit you would bring

Before choosing a tool, assemble the assets any of them could draw on:

• Character reference: a few clean stills of Elena from consistent angles to lock identity and wardrobe.
• Product reference: high-resolution images of the ring from multiple angles, plus the packaging.
• UI reference: a pixel-accurate export of the app dashboard so on-screen text and layout stay legible.
• Motion reference (optional): a short clip of the desired walking-track camera move, for tools that accept video-to-video guidance.
• Audio reference: a scratch music bed, a recording of the dialogue line, and the sonic logo, for tools that accept audio references or generate sound in the same pass.
• Shot list: a one-page plan with the three time blocks and the exact CTA wording.

How completely each tool can ingest that kit is the first real point of difference, covered below.

Creative requirements at a glance

The matrix below maps each requirement in the brief to how the three options are documented to handle it. It is a planning summary, not a scorecard, the detail behind each cell follows in the sections after it.

Beat by beat: what each stretch of the timeline demands

0–5 seconds — the cold open

The opening has to do two jobs inside a very short window: establish mood and put the product on screen. It opens on the ring resting on a nightstand in sunrise light, a macro insert before a soft cut to Elena stirring. The demands are a clean macro product rendering and an identity-consistent first frame of the character.

Any of the three can, in principle, produce a five-second beat; what matters is whether that beat can live inside the same generation as the rest of the spot, so the light and grade match, or has to be produced separately and matched in the edit.

5–15 seconds — the workhorse

The middle stretch is where three hard things stack up at once: Elena on the street in a handheld tracking shot, the ring visible on her finger, a quick glance at the app dashboard, and the spoken line.

That means sustained character consistency across a moving shot, a legible UI insert, and dialogue with believable lip sync in the same ten seconds. It is the most demanding block in the brief and the best stress test of each model’s control.

15–30 seconds — the payoff

The final stretch runs longest. A daytime studio beat and an evening return home carry the “recovery and focus” idea, then a slow orbit hero shot of the ring resolves into the end card, CTA, sonic logo, and the musical button.

The requirement is continuity across a location and time-of-day change, a controlled hero move, and precise timing so the audio button lands exactly with the CTA. Timing precision is the quiet difficulty here, the music, the last frame, and the CTA all have to hit together.

References: how each option ingests the kit

Seedance 2.5 has the widest documented reference capacity of the three: up to fifty assets in a single generation, split as up to thirty images, ten video clips, and ten audio files. For this brief that means Elena’s stills, the ring and packaging images, the UI export, an optional motion clip, and the scratch audio can all be handed to one generation. It also documents storyboard, keyframe, white-model previs, and green-screen references for staging control.

MiniMax H3 is natively multimodal across text, image, video, and audio, and its omni-reference system is documented in MiniMax’s own materials as accepting up to nine reference images, three video clips, and three audio clips at once. That is ample for a single shot, a product image, a motion clip, a voice track, but it is a smaller stack than Seedance’s, which matters when you want one generation to hold the character, the product, the packaging, and the UI at the same time.

Its particular strength in the reference department is text and brand fidelity: it is positioned to preserve logos, typography, and product cues more reliably, which is exactly what the app-UI insert and the end card need.

Wan 3.0 (preview) is presented with the broadest reference concept on paper, an “omni” brief that groups image, text, video, and audio and adds documents and webpages as structured context, which in principle would let you feed the product brief or the live product page straight into the plan.

The caveat is that this is a Topview workflow preview: the reference set, like the rest of the Wan 3.0 specification, is projected rather than confirmed, so the practical rule is to verify what the live generator actually accepts before building a kit around it.

Dialogue and audio

The brief needs one clean spoken line with convincing lip sync, a music bed, and a sonic logo, all timed to the cut.

Seedance 2.5 generates video and audio together in a single pass and documents multilingual dialogue with improved lip synchronisation and more reliable subtitles, plus audio references and time-coded sound direction. For a single-line delivery with a music button, that is a strong fit: the dialogue, the bed, and the sound can be directed within the same generation and, via second-level timestamps, placed against the picture.

MiniMax H3 also produces native synchronised audio, stereo, in the same pass as the picture, across a set of stably supported languages. Its audio quality and one-pass dialogue are genuine strengths; the constraint is not the audio but the length ceiling, which shapes how much of the spot any single H3 generation can cover.

Wan 3.0 (preview) treats voice, music, and sound as part of the creative brief, but Topview is explicit that whether native audio or audio-reference input is actually available depends on the live model mode. For a deliverable with a timed music button, that uncertainty means audio timing should be confirmed in the product, not assumed from the preview page.

The product shot

Two things define a good product beat for the Loop: a crisp macro that holds the ring’s material and geometry, and a legible app dashboard where the on-screen text does not smear.

For the macro, MiniMax H3’s native 2K output is attractive, more resolution in the hero insert, and its brand-fidelity positioning suits the packaging and logo moments. For the dashboard, both H3’s text fidelity and Wan 3.0’s documented UI, text, and animation direction are aimed squarely at readable interface renders, though the Wan capability is preview-stage and, like all AI UI rendering, needs frame-by-frame checking for distorted labels.

Seedance 2.5 can hold the product across the whole take thanks to its reference stack and single-pass continuity, which is the advantage when the same ring has to look identical in the macro, on the finger, and in the orbit. In practice you would likely let Seedance carry product continuity across the spot and consider H3 for a maximally sharp standalone hero macro if 2K matters for the placement.

Revisions

Real campaigns are revised. The client will want the third beat re-timed, or the line re-read, or the packaging swapped without re-rolling the entire film.

This is where Seedance 2.5’s documented editing is the differentiator. It supports local segment editing, revising a selected section while keeping the rest more stable, alongside timestamp-level control, green-screen replacement, camera-perspective re-editing, and reference-based editing, and it can extend or continue a strong clip. For a 30-second spot that will go through rounds, being able to re-do only the 15–30s block, or swap only the packaging shot, is a major planning advantage.

MiniMax H3 is documented for precise, instruction-led editing and video-to-video motion transfer, so targeted changes within its 15-second window are well supported; the ceiling simply means revisions are scoped to segments rather than to a single 30-second master.

Wan 3.0 (preview) describes reference-led consistency and shot-level direction, which would support revision in principle, but the preview status means the actual editing controls should be confirmed live before you promise a client round-based changes on it.

Horizontal, vertical, and square delivery

The brief needs a 16:9 master plus 9:16 and 1:1 versions.

Wan 3.0’s preview page documents the widest aspect-ratio set of the three — 16:9, 9:16, 1:1, 4:3, and 3:4, which, if the live generator supports it, would be convenient for multi-format delivery from one brief. Again, projected, so confirm.

Seedance 2.5’s Topview generator shows a 16:9 example prominently; other ratios and a native vertical render should be confirmed in the live generator, and where a native vertical pass is not exposed, the practical fallback is to frame the horizontal master with vertical safe-areas in mind and reframe in the edit. MiniMax H3 lists an “adaptive” aspect ratio on Topview, which suggests flexibility but is also worth confirming per placement.

The honest planning position: do not assume any of the three will hand you all three ratios natively. Design the horizontal master so it can be reframed, and treat native vertical and square renders as a confirm-in-product step.

How each option would approach the whole brief

Seedance 2.5 — the natural spine

It is the only one of the three that documents a native 4–30-second single continuous generation, so the entire spot can be planned as one coherent take with the light, grade, and character held across the full thirty seconds.

Add second-level timestamps for the 0–5 / 5–15 / 15–30 structure, up to fifty references for identity and product, in-pass audio for the line and the music, and local segment editing for revisions, and Seedance can carry the brief end to end. Its main watch-item is output resolution: ByteDance’s launch materials describe native 4K and 10-bit colour, while Topview’s generator currently exposes up to 1080p, fine for social and product-page use, but worth confirming if a higher master is required.

MiniMax H3 — the specialist

Its 15-second ceiling means it cannot hold the whole spot in one generation; a 30-second brief on H3 alone must be built as at least two segments and assembled. What it offers instead is quality where quality is most visible: native 2K, native stereo audio, strong instruction-following, and text and brand fidelity that suit the app-UI insert, the packaging, and the end card.

Two further planning notes apply. First, H3’s open weights allow private, self-hosted deployment, appealing if sensitive brand assets should not leave your infrastructure, but its community licence is documented as excluding local deployment in the United States, European Union, United Kingdom, and South Korea, which matters for teams operating there. Second, on a hosted platform such as Topview those licensing mechanics are abstracted away, but commercial-use terms should still be checked per export.

Wan 3.0 — the evaluation lane

Treat it as promising but provisional. Topview is unusually candid here: the page states that, as of its August 12, 2026 review, Alibaba Cloud’s public model lifecycle documented Wan 2.7 as the latest listed Wan video release, and that Wan 3.0’s durations, modes, and other settings shown are workflow previews rather than confirmed specifications.

Third-party trackers separately describe a wan3.0-video identifier appearing on Alibaba Cloud Model Studio for beta or invite testing in early-to-mid August 2026, but the specifications remain unconfirmed and, for this brief, no firsthand testing was performed. The sensible plan is to prototype the omni-reference brief and the multi-format delivery in Wan 3.0 if the live generator supports them, while keeping Seedance 2.5 or the officially documented Wan 2.7  as the fallback for anything that has to ship.

Where the workflow has to split

No single one of the three is a clean fit for every requirement at once, so the realistic plan is a division of labour rather than a single button:

• Spine on Seedance 2.5: the continuous 30-second build, character continuity, in-pass dialogue and music, and round-based revisions via local segment editing.
• Specialist inserts from MiniMax H3 where its 2K and text/brand fidelity are most visible, a standalone hero product macro, or a maximally clean end-card and UI beat  accepting that these are 15-second-or-shorter segments cut into the Seedance spine.
• Wan 3.0 as an evaluation and multi-format lane only, used to test the omni-reference approach and the wider aspect-ratio set, never as the sole source for a shipping deliverable while its specifications are unconfirmed.

The reason to name the split explicitly is that the alternative, pretending one model natively satisfies a recurring character, multiple locations, dialogue, a legible UI, in-pass music, round-based revisions, and three delivery ratios all at once, would misrepresent what these tools currently do.

Constraints to plan around

• Length ceilings differ sharply: Seedance 2.5 documents up to 30s in one pass; MiniMax H3 tops out around 15s; Wan 3.0’s 30s is projected, not confirmed.
• Resolution is platform-dependent: H3 is native 2K; Seedance is described as native 4K at the model level but exposed up to 1080p on Topview; Wan 3.0 (preview) lists 1080p and 720p.
• Audio availability is certain for Seedance and H3 (in-pass) but preview-stage for Wan 3.0.
• Reference capacity is widest on Seedance (up to 50), moderate on H3 (about 9 images, 3 video, 3 audio), and broad-but-projected on Wan 3.0.
• Licensing and region: H3’s open-weight local deployment is restricted in several major markets; hosted use abstracts this but still carries per-export commercial terms across all three.
• UI and text rendering always needs frame-by-frame review regardless of model — small on-screen text is where AI video most often breaks.
• Delivery ratios should be confirmed in the product; design the horizontal master to be reframeable rather than assuming native vertical or square output.

Documented capability snapshot

A compact reference for the specifications discussed above. Figures reflect documentation current at the time of writing and can change per platform and account.

Final selection

If the goal is one signed-off 30-second launch film, the plan writes itself around Seedance 2.5 as the anchor: it is the only option of the three that documents a native single-pass thirty seconds, timestamped beat control, a fifty-asset reference stack, in-pass audio, and segment-level revision, the exact combination a round-based client deliverable needs.

MiniMax H3 earns a defined specialist role for the shots where its native 2K and brand and text fidelity are most visible, on the understanding that its fifteen-second ceiling makes it a contributor of inserts rather than the carrier of the whole spot. Wan 3.0 stays in the evaluation lane: genuinely interesting on paper, wider on reference concept and aspect ratios, but preview-stage and untested here, so it informs the plan without being trusted to deliver it.

The honest headline is not “one model wins” but “one model anchors, one specialises, and one is worth watching”  and the workflow is built accordingly.

A note on sourcing, status, and Topview-specific tools

Capability claims for Seedance 2.5 and MiniMax H3 are drawn from each model’s documented behaviour, including the vendors’ own release materials and the corresponding Topview generator pages.

All Wan 3.0 specifics are preview-stage: Topview presents them as a workflow preview and, as of its stated August 12, 2026 review, pointed to Alibaba’s public catalog listing Wan 2.7 as the latest documented Wan release. No campaign was generated and no model was tested first-hand; durations, resolutions, and reference limits reflect documentation current at the time of writing and can change per platform and account.

Topview-specific tools (labeled). Several capabilities that sit around these models on Topview are platform features rather than model-native ones. The 3D Director Console and white-model previs, the Conversational Canvas editing surface, and Douyin-link recreation are Topview workflow tools built around the underlying models.

Likewise, video extension and the specific reference and export ceilings shown are properties of the Topview generator as configured, not necessarily of the base model in every deployment. Where this article relies on those features, it is describing the Topview workflow, not an inherent property of Seedance, MiniMax, or Wan.

—

This content is brought to you by AK Newswire

istockphoto